Prediction of Compression Coefficients Based on Machine Learning: A Case of Offshore Wind Farm Site

Machine learning methods have a wide range of applications, including predicting soil compression coefficients for offshore wind power projects. This study compared three machine learning methods (Extreme Gradient Boosting, Random Forest, and Bayes) with traditional empirical formulations. Results revealed distinct spatial distribution characteristics within the soil layers, with the upper clay layer exhibiting high water content, void ratio, mobility, and compressibility, the middle clay layer showing medium compressibility, and the bottom layer consisting of sandy soil. All three models outperformed the traditional empirical formulation, with Extreme Gradient Boosting demonstrating the best performance (MAPE = 16.9, MAE = 0.10, MSE = 0.024, RMSE = 0.16, R² = 0.91). Parameters influencing the compressibility coefficient were categorized into three groups based on SHAP values, with density, water content, and liquid index having the greatest impact. The compression coefficient showed negative correlations with density, depth, dry density, and positive correlations with water content, liquid index, and pore ratio. The factors influencing soil compression coefficient prediction can be attributed to internal soil structure and external conditions. Additionally, a new formula with an R² value of 0.95 was proposed for predicting compression coefficient based on density, water content, and liquid index, offering practical guidance for engineering applications.